An Introduction to NVIDIA Jetson

Artificial intelligence is moving beyond centralized data centers and into the physical world, where robots, machine vision systems, and autonomous machines must process growing volumes of sensor data and make decisions in real time.

This shift is being driven in large part by the GPU. Originally designed for graphics rendering, GPUs have evolved into powerful parallel processors capable of executing large numbers of calculations simultaneously. This makes them especially well-suited for AI inference, computer vision, robotics, and other data-intensive workloads where CPUs alone may not provide enough parallel processing performance.

However, conventional discrete GPUs can introduce challenges for embedded and industrial deployments. High-performance GPUs often require significant power, generate substantial heat, and depend on active cooling, making them difficult to integrate into compact systems operating in space-, power-, or thermally constrained environments.

NVIDIA Jetson addresses these challenges by combining power-efficient GPU-accelerated computing, CPU processing, memory, and high-speed interfaces into compact embedded AI platforms supported by NVIDIA's CUDA and accelerated software ecosystem.

Today, the portfolio is led by two major families:

• NVIDIA Jetson Orin for scalable edge AI, machine vision, and robotics.
• NVIDIA Jetson Thor for advanced Physical AI, multimodal AI, and next-generation robotics.

What is NVIDIA Jetson?

NVIDIA Jetson is a family of System-on-Modules, or SoMs, designed to bring GPU-accelerated AI computing directly to edge devices.

Instead of building a system around separate CPU, GPU, and memory components, a Jetson module integrates key computing resources into a compact embedded platform.

Depending on the Jetson family, these resources can include: 

• NVIDIA GPU and Tensor Cores
• Arm-based CPU
• High-speed system memory
• Hardware video processing
• Camera and sensor interfaces
• PCIe and high-speed networking
• Dedicated AI and vision accelerators

The result is a compact computing architecture that can be integrated into robots, autonomous vehicles, smart cameras, medical equipment, industrial machines, and other embedded AI systems. 

Jetson is also more than hardware. NVIDIA combines its modules with JetPack SDK, CUDA, TensorRT, cuDNN, DeepStream, Isaac, Holoscan, and other software tools used to develop and deploy accelerated AI applications. 

NVIDIA describes Jetson as its platform for AI at the edge, combining power-efficient embedded compute with its AI software stack for advanced robotics and autonomous products.

The Evolution of NVIDIA Jetson

Earlier Jetson generations such as Jetson Nano, TX2, and Xavier helped establish GPU-accelerated computing for embedded AI, paving the way for today’s higher-performance platforms.

Jetson Orin expanded the portfolio for AI inference, robotics, computer vision, and generative AI within embedded power envelopes, while Jetson Thor takes the next architectural step with NVIDIA Blackwell GPU technology, greater memory capacity, and significantly more AI compute for complex Physical AI and multimodal workloads.

Rather than replacing Orin, Thor extends how far the Jetson platform can scale.

NVIDIA Jetson Orin

The NVIDIA Jetson Orin family is designed to scale AI performance across a wide range of edge deployments.

Built on NVIDIA Ampere architecture GPUs and Arm® Cortex®-A78AE CPUs, the Orin portfolio is available in three main tiers—Orin Nano, Orin NX, and AGX Orin—with different levels of CPU performance, AI compute, memory capacity, and power consumption.

The table below compares the key specifications and typical application fit for each Jetson Orin tier.

NVIDIA Jetson Orin Super Mode

Super Mode allows supported Jetson Orin modules to unlock higher AI performance by increasing GPU operating frequencies and enabling higher power configurations. Rather than requiring a new module, it extends the performance available from existing Orin hardware through supported JetPack software and power modes.

With Super Mode, Jetson Orin Nano reaches up to 67 TOPS, while Jetson Orin NX reaches up to 157 TOPS. JetPack 7.2 also introduced Super Mode for Jetson AGX Orin 32GB, increasing peak AI performance from 200 TOPS to 241 TOPS.

The additional performance can benefit workloads such as multi-camera vision, robotics, generative AI, and multimodal inference. However, higher performance also increases power consumption and thermal requirements, so system designers must ensure the platform’s cooling and power delivery are designed to support the selected Super Mode configuration.

NVIDIA Jetson Thor: The Next Step for Physical AI

As edge AI becomes more sophisticated, the workload is changing. 

Traditional machine vision may need to answer, “What object is in front of me?” A Physical AI system may need to understand its surroundings, reason about what is happening, determine what to do next, and physically respond in real time. 

Supporting these workloads requires greater AI compute, memory bandwidth, and processing capacity. This is where NVIDIA Jetson Thor extends the Jetson platform. 

Built on NVIDIA Blackwell architecture GPUs with fifth-generation Tensor Cores and Arm-based processors, the Jetson Thor family is designed for advanced robotics, multimodal AI, visual AI agents, and Physical AI applications. 

The Thor family now scales across four modules, from the recently announced T2000 and T3000 to the higher-performance T4000 and T5000.

The Jetson Thor lineup gives developers a scalable path for deploying Blackwell-based AI at the edge, from more compact robotics and visual AI workloads with T2000 and T3000 to advanced Physical AI and multimodal applications with T4000 and T5000. As compute, memory, and power scale across the family, system designers can better match the platform to the complexity of the workload rather than defaulting to the highest-performance option. This makes Jetson Thor suitable for a broader range of next-generation robotics, autonomous systems, sensor fusion, and real-time AI applications.

Jetson Orin vs. Jetson Thor

Jetson Orin and Jetson Thor address different levels of edge AI performance. Jetson Orin focuses on scalable, power-efficient AI inference for machine vision, robotics, and autonomous systems, while Jetson Thor targets more demanding Physical AI workloads that require larger models, more memory, and advanced multimodal processing.

The choice between Orin and Thor should therefore start with the workload rather than maximum AI performance. For machine vision, video analytics, AMRs, and many industrial AI applications, Jetson Orin provides a strong balance of compute performance, power efficiency, size, and memory. Jetson Thor becomes more relevant when systems must process larger multimodal models, combine multiple high-bandwidth sensor streams, or perform more complex sense → understand → reason → act workflows in real time.

Learn more about Jetson Orin vs Jetson Thor.

Note: Orin and Thor use different AI performance metrics, so their headline compute figures should not be compared directly.

Choosing the Right AI Compute Platform

NVIDIA Jetson does not replace discrete GPUs. Instead, it gives OEMs and system integrators another way to deploy GPU-accelerated AI based on the performance, power, size, and software requirements of the application.

Different AI architectures are better suited for different workloads:

• Discrete GPUs: Best for applications that require maximum compute performance, large models, high memory capacity, or flexible PCIe expansion. They are well suited for AI training, high-performance inference, and workloads where power and thermal constraints are less restrictive.
  
  
• NVIDIA Jetson: Designed for AI at the edge, where compact size, power efficiency, low latency, and local processing are priorities. Jetson is particularly well suited for machine vision, robotics, autonomous systems, and Physical AI.
  
  
• Dedicated AI accelerators: NPUs and other purpose-built AI accelerators offer another option for efficient edge inference. These accelerators can be integrated into existing x86 or Arm-based systems to add AI performance without requiring a complete change in system architecture, making them useful for power-conscious vision and inference workloads. 

There is no single architecture that fits every AI deployment. The right choice depends on factors such as model complexity, performance requirements, power budget, system size, software compatibility, and scalability.

To explore these architectures in more detail, read our complete overview of CPUs, GPUs, and AI accelerators.

Conclusion 

As AI moves closer to where data is generated, system designers need computing platforms that balance AI performance, power consumption, system size, and deployment requirements. NVIDIA Jetson addresses this need by bringing GPU-accelerated AI into compact embedded platforms designed for real-time edge processing. 

Within the portfolio, Jetson Orin provides scalable performance for machine vision, robotics, and autonomous systems, while Jetson Thor extends the platform toward more demanding Physical AI, multimodal AI, and advanced robotics workloads. Choosing between Jetson, discrete GPUs, or other AI accelerators ultimately depends on the workload, power budget, software requirements, and level of scalability required by the application. 

As edge AI continues to evolve from basic inference toward systems that can sense, understand, reason, and act, NVIDIA Jetson provides a scalable foundation for bringing increasingly capable AI into real-world machines and industrial environments. 

Premio’s latest lineup of Jetson Orin AI Edge Computers taps into the design of Jetson Orin to offer a new family of rugged, fanless edge AI industrial computers. The JCO Series is Premio’s first line utilizing ARM-based architecture, offering three scalable models from entry-level and mid-range, JCO-1000-ORN Series and JCO-3000-ORN Series powered by Jetson Orin Nano and Orin NX to the high-performance JCO-6000-ORN Series powered by Jetson AGX Orin.

Premio is also expanding its lineup with the upcoming WCO-6000-THR Series, powered by NVIDIA Jetson AGX Thor and planned for availability in 2027, extending the portfolio toward more demanding Physical AI, multimodal processing, and advanced robotics workloads.

This blog was originally published on April 5, 2024, and has been updated to reflect the latest product and certification information.

jco-series